S197
Clinical - Biomarkers of clinical response
ESTRO 2026
prostate cancer using urine samples. The RFDD technique enhances analytical precision and diagnostic accuracy, showing promise for clinical application and community screening. Keywords: prostate cancer, mass spectrometry, diagnosis Digital Poster 4286 Utilizing Machine Learning on MR-Linac daily images as biomarkers for GBM patients’ response prediction in Adaptive RT: preliminary results Edoardo Salmeri 1,2 , Roberto Pellegrini 1 , Guido Baroni 2 , Chiara Paganelli 2 , Arjun Sahgal 3 , Chia-Lin Eric Tseng 3 , Angus Lau 3 1 Medical Affairs and Clinical Research, Elekta, Stockholm, Sweden. 2 Department of Electronics, Information and Bioengineering (DEIB), Politecnico di Milano, Milan, Italy. 3 Department of Radiation Oncology, Sunnybrook Health Sciences Centre, Toronto, Canada Purpose/Objective: Glioblastoma (GBM) is an aggressive grade IV brain tumor, representing 14% of all brain tumors, with a median survival of 15 months [1]. This study aims to leverage daily diagnostic MRI acquired directly on the 1.5T Elekta MR-Linac (Unity) during treatment sessions to predict radiotherapy (RT) outcomes. By capturing longitudinal imaging features from each treatment day, we aim to model both spatial tumor characteristics and temporal evolution. The goal is developing a deep learning framework capable of early prediction of treatment response, enabling adaptive, patient-specific RT decisions. Material/Methods: We analyzed a retrospective cohort of 68 patients with GBM treated on Unity. For each patient, longitudinal MRI data were collected, including daily T1 and FLAIR sequences (9–30 days per patient), planning CT, and structure set. MR images were skull-stripped [2], resampled, and bias-corrected using N4 denoising. MRI were rigidly registered to the patient’s planning CT [3]. Initial planning target volume (PTV) masks were used to crop MR volumes. Each sequence was independently z-score normalized per patient (Fig.1). Longitudinal T1 and FLAIR volumes were used as separate channels. A 3-month follow-up label was available for each patient: Stable Disease, Progression or Partial Response.
A 3D convolutional neural network (CNN) was employed as a feature extractor, with the final layer using adaptive average pooling to produce a fixed-size feature vector. The CNN output for each day was fed into a Long Short-Term Memory (LSTM) network to model temporal tumor evolution. The LSTM output at the last time point passed through a fully connected classifier to predict one of the response classes. This architecture allows extraction of spatial features from 3D MR volumes, capturing longitudinal dynamics. Results: The CNN+LSTM model trained on longitudinal T1 and FLAIR MRIs (80%) achieved training accuracy up to 87% (macro F1 0.86) over 30 epochs. Validation accuracy on the remaining 20% dataset ranged 21–57% (macro F1 0.12–0.41) (Fig.2). These results suggest the model captures temporal and spatial patterns but requires larger, balanced datasets for generalization.
Conclusion: We preliminary evaluated the feasibility of an
integrated CNN+LSTM model using longitudinal MRIs to predict treatment response in GBM. Performance is limited by dataset size, class imbalance, and low generalization. Future work will focus on integrating handcrafted radiomics features alongside deep learning outputs [4], addressing class imbalance, and improving model interpretability to enhance predictive accuracy and clinical utility. References: [1] Stupp R et al, 2005, N Engl J Med, https:// doi. org/ 10.1056/NEJMoa043330[2] Isensee et al., 2021, Nature methods, https://doi.org/10.1038/s41592-020-01008- z[3] Park H et al., 2024, PLoS One, doi: 10.1371/journal.pone.0299366[4] Tikhonov, D et al.,
2025, arXiv, doi:10.48550/arXiv.2509.06511 Keywords: MR-Linac, Longitudinal, Response
Made with FlippingBook - Share PDF online